Seite - 180 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

energies Article Short-TermLoadForecastingofNaturalGaswith DeepNeuralNetworkRegression† GregoryD.Merkel,RichardJ.Povinelli * ID andRonaldH.Brown OpusCollegeofEngineering,MarquetteUniversity,Milwaukee,WI53233,USA; gregory.merkel@marquette.edu(G.D.M.); ronald.brown@marquette.edu(R.H.B.) * Correspondence: richard.povinelli@marquette.edu;Tel.:+1-414-288-7088 † Thiswork isanextensionof thepaper“Deepneuralnetworkregressionforshort-termloadforecastingof naturalgas”presentedat the InternationalSymposiumonForecasting,17–20 June2015,Cairns,Australia, andispublishedin theirproceedings. Received: 29 June2018;Accepted: 1August2018;Published: 2August2018 Abstract: Deep neural networks are proposed for short-term natural gas load forecasting. Deeplearninghasproventobeapowerful tool formanyclassificationproblemsseeingsignificant use inmachine learningfields suchas image recognitionand speechprocessing. Weprovide an overviewof natural gas forecasting. Next, the deep learningmethod, contrastive divergence is explained.Wecompareourproposeddeepneuralnetworkmethodtoa linear regressionmodeland atraditionalartificialneuralnetworkon62operatingareas,eachofwhichhasat least10yearsofdata. Theproposeddeepnetworkoutperformstraditionalartificialneuralnetworksby9.83%weighted meanabsolutepercenterror (WMAPE). Keywords: short termloadforecasting;artificialneuralnetworks;deep learning;naturalgas 1. Introduction Thismanuscriptpresentsanoveldeepneuralnetwork(DNN)approachtoforecastingnatural gas load. We compare our newmethod to three approaches—a state-of-the-art linear regression algorithmandtwoshallowartificialneuralnetworks(ANN).Wecompareouralgorithmon62datasets representingmany areas of theU.S. Eachdataset consists of 10 years of trainingdata and 1 year of testingdata. Our newapproach outperforms each of the existing approaches. The remainder of the introduction overviews the natural gas industry and the need for accurate natural gas demandforecasts. Thenaturalgas industryconsistsof threemainparts;productionandprocessing, transmission andstorage,anddistribution[1]. Likemanyfossil fuels,naturalgas (methane) is foundunderground, usuallynearorwithpocketsofpetroleum.Naturalgasisacommonbyproductofdrillingforpetroleum. Whennaturalgas iscaptured, it isprocessedtoremovehigheralkanessuchaspropaneandbutane, whichproducemoreenergywhenburned.After thenaturalgashasbeenprocessed, it is transported viapipelinesdirectly to localdistributioncompanies (LDCs)orstoredeitheras liquidnaturalgas in tanksorbackunderground inaquifersor salt caverns. Thenaturalgas ispurchasedbyLDCswho providenaturalgas toresidential, commercial, andindustrial consumers. Subsetsof thecustomersof LDCsorganizedbygeographyormunicipalityarereferredtoasoperatingareas.Operatingareasare definedbythe individualLDCsandcanbeas largeasastateorassmallasa fewtowns. Theamount ofnatural gasusedoften is referred toas the loadand ismeasured indekatherms (Dth),which is approximately theamountofenergy in1000cubic feetofnaturalgas. ForLDCs, thereare severalusesofnaturalgas, but theprimaryuse is forheatinghomesand businessbuildings,which is calledheatload. Heatloadchangesbasedon theoutside temperature. Energies2018,11, 2008;doi:10.3390/en11082008 www.mdpi.com/journal/energies180